YAGO Knowledge Base

Inputs

Precomputed Values

Experiment Results

Inputs

Predefined input domains.

actors , albums , books , cities , countries , singers , songs , writers

 

New input domains.

actors id from themoviedb , singers id from deezer , singers id from discogs , singers id from echonest , singers id from last_fm , singers id from music brainz , singers id from musixmatch , writers id from isbndb , writers id from library thing

Precomputed Values

Functionality of relations.

Relations with functionality > 0.9.

Relations with functionality < 0.1.

Experiment Results

Paths Alignment

In the following table are presented the precision and the recall for each service (averaged over the APIs that provide the service). We ran experiments for both alignment strategies: the overlapping and the subsumption strategy. For the overlapping strategy, we show also the results of the method containing data cycles (WithCycles). For the subsumption, we checked whether the results of the KB path are subsumed by the results of the XML path (KB → W S), and vice-versa (W S → KB). Click on the table for more details.

Path Alignment

 

The results of our path alignment algorithm and the goldset, for each Web service are in the folder:

Path Alignment (Overlapping) - Dropbox Folder.

 

Path Alignment (Association Rule Mining ) - Dropbox Folder.

 

Classes and Relations Alignment

We tested the class and relation alignment that keeps duplicates (With dup.), and the alignment that removes duplicates (DORIS). Our evaluation shows that duplicates can be safely removed: the precision in the relations increases drastically while the recall is unchanged in a majority of cases and relations.

Moreover we compared our system with the state-of-the-art alignment approach PARIS. We ran PARIS with exactly the same data as DORIS. PARIS computes a confidence score for each alignment. We used a threshold of 0.6 on this score to determine the final output, which was the value for which PARIS performed best. The results of this approach in relation alignment are shown in the following table under the columns "PARIS". As we see, DORIS outperforms PARIS by a huge margin. DORIS achieves near-perfect alignment, while the precision and recall for PARIS are more in the area of 30%-60%. The reason is that PARIS performs very well when the schemas of the two ontologies are similar. This is not the case at all in our problem. Moreover, PARIS cannot discover complex relations alignments like hasGender.label, while DORIS discovered them.

Click on the table for more details.

Classes-Relations Alignment

 

The results of our class and relation alignment algorithm and the goldset, for each Web service are in the folder:

Class and Relation Alignment - Dropbox Folder.

 

Views and Transformation Functions

The results of the views definition using class and relation alignment results are in the folder:

Views - Dropbox Folder.

 

 

The xslt code is produced by using the class & relation alignment result of each function are in the following folder:

XSLT - Dropbox Folder.

You can test the result on the online XSLT Editor of w3schools here by using the code that we produce and a call result of a Web service.

 

Input/Output Dependencies

Following table shows the precision and the recall for several variants of the algorithm: S1 and S2 implement Step 1 and Step 2 respectively. S2-3 implements Steps 2 and 3 together. We find that Step 1 sometimes excludes good answers, while S2-3 achieves a precision that is consistently over 0.8. All strategies have a very high recall, which is almost perfect for S2 and S2-3.

IO Dependencies